The definitive academic writing standard, stylometric calibration, and automated thesis formatting engine for higher education and peer-reviewed research. Refines synthetic drafts into rigorous, empirical, publication-grade academic prose and compiles them directly into 100% compliant Microsoft Word (
.docx) documents.
Most commercial paraphrasers (QuillBot, Undetectable.ai, StealthGPT) fail modern institutional AI detectors (Turnitin AI, GPTZero, ZeroGPT, Originality.ai) because they merely perform synonym substitution over preserved AI syntactic skeletons.
Modern neural AI detectors do not read semantic meaning; they evaluate statistical text dynamics:
- Perplexity (Token Predictability): AI models predictably select tokens with the highest statistical likelihood (Top-K most predictable tokens).
-
Burstiness (Sentence Length & Cadence Variation): AI generates monotonous, metronomic sentence structures (averaging 14–20 words per sentence, standard deviation ratio
$\sigma/\mu < 0.30$ ). Authentic human academic writing exhibits extreme dynamic range: punchy 3–6 word staccato observations contrasted with 35–45 word compound technical clauses ($\sigma/\mu \ge 0.55$ ). -
Syntactic Clichés & Stacking: Monotonous nested clauses ("is built by combining X to do Y and Z in order to W"), unprovoked em dashes (
—), and hollow textbook nominalizations (crucial, comprehensive, holistic, delve, foster, paradigm shift).
graph TD
A[Raw AI Draft / Thesis Section] --> B[Phase 1: Lexical Purge]
B -->|Strip AI-isms, Clichés & Em Dashes| C[Phase 2: Burstiness & Cadence Oscillation]
C -->|Sentence Length Variation sigma/mu >= 0.55| D[Phase 3: Methodological Hardening]
D -->|Inject Grounded Empirical Parameters| E[Phase 4: Universal DOCX Compilation]
E --> F[Publication-Ready Word Document .docx]
- Strips all Tier-1 and Tier-2 AI clichés across both English (delve, robust, pivotal, intricate, multifaceted, testament to, foster, leverage) and Indonesian (berperan krusial, sangat penting, menelisik lebih dalam, merajut, paradigma baru, secara holistik).
- Hard-strips em dashes (
—), converting them into commas, parentheses, semicolons, or distinct sentences. - Eliminates mid-paragraph bolding (
**text**) reserved exclusively for headings and table labels.
- Forces sentence length variance to meet or exceed a Coefficient of Variation
$\sigma / \mu \ge 0.55$ . - Injects punchy 3–6 word factual anchors in every paragraph: "Field conditions were erratic.", "Gradients vanished immediately.", "Memory overflowed instantly."
- Pairs these anchors with data-dense 30–45 word compound observations containing real operational figures and comparative clauses.
- Purges dictionary and textbook definition dumps from methodology sections ("Quantitative research is defined as...").
- Replaces abstractions with verifiable empirical procedures: specific instruments, error margins, hardware specifications, sampling frames, and outlier treatment criteria.
- Assembles clean, publication-grade Word documents via native OpenXML:
- Paper Format: Standard A4 (80 gsm), Margins: Left 4.0 cm (binding margin), Top 2.5 cm, Right 2.5 cm, Bottom 2.5 cm.
- Typography: Times New Roman 12pt, 1.5 line spacing, Justified alignment, 1.0 cm first-line indentation.
- Color Sanitation: 100% Pure Black (
#000000), purging web clipboard gray bleed (#374151). - Scientific Tables: APA-style 3-line open tables (three horizontal borders, zero vertical column lines).
| Element | Specification & Action | Target |
|---|---|---|
Em Dash (— / --) |
Hard Strip. Forbidden in formal academic prose. Replace with commas or split sentences. | 0 occurrences |
In-Text Bold (**text**) |
Hard Strip. Mid-paragraph bolding is prohibited. Bold is strictly reserved for Headings and Table labels. | 0 occurrences |
| Textbook Definition Dumps | Hard Strip. Delete generic definitions of standard methodologies. Replace with operational parameters. | 0 dumps |
| Font Color Bleed | Sanitize. Enforce 100% Pure Black (#000000). Strip web clipboard off-black colors. |
100% #000000 |
| Symmetrical Bullet Lists | Auto-convert. Transform symmetrical bullet points into coherent narrative prose paragraphs. | Flowing Prose |
| Non-Standard Technical Terms | Enforce Italics. Italicize all Latin, foreign, or non-dictionary technical terminology. | Strict Italics |
Designed with Zero Friction: run directly via uv without manual virtual environment management.
# Clone the repository
git clone https://github.com/demusraph/academic-writing-engine.git
cd academic-writing-engine
# 1. Run live verification against the public ZeroGPT API (Target: 0.0% AI)
uv run scripts/verify_zerogpt.py draft.txt
# 2. Run local statistical Burstiness & Perplexity audit (100% offline, zero-auth)
uv run scripts/verify_gptzero.py draft.txt --audit
# 3. Run live public allowance check on Originality.ai (zero API key required)
uv run scripts/verify_originality.py draft.txt --live
# 4. Compile Markdown/Text draft directly into an official APA/IEEE Word (.docx) document
uv run --with python-docx scripts/export_academic_docx.py draft.txt thesis_final.docxgit clone https://github.com/demusraph/academic-writing-engine.git
cd academic-writing-engine
pip install -r requirements.txt
# Compile DOCX document
python scripts/export_academic_docx.py draft.txt thesis_final.docxThis repository includes standalone verification tools that do not require private user tokens, cookies, or browser inspection:
Queries the official live ZeroGPT public endpoint directly (api.zerogpt.com/api/detect/detectText):
python scripts/verify_zerogpt.py draft.txt
# Output:
# Status: SUCCESS | AI Score: 0.0% | Feedback: Your Text is Human Written- Supports
--auto-cleanto iteratively calibrate flagged sentences down to 0.0% AI.
Runs 100% locally on your machine without external network dependencies:
- Measures sentence length variance (
$\sigma/\mu$ ) to ensure Burstiness exceeds 0.55. - Evaluates sentence-level Perplexity proxies and detects AI syntactic markers.
python scripts/verify_gptzero.py draft.txt --auditQueries the public free allowance endpoint (api/v2-tools/free-tools/ai-allowance):
python scripts/verify_originality.py draft.txt --live --threshold 15Technical Note: Originality.ai's public free scanner runs an English-only model (
en). When evaluating non-English text, ensure terminology includes grounded hardware and technical descriptors to avoid cross-language BPE tokenization penalties.
To ensure the engine does not overfit to a single domain, we include 130 Master Benchmark Test Cases in benchmarks/academic_130_benchmark.json:
- 105 STEM Cases: Software Engineering, Cybersecurity, Industrial IoT, Computer Vision, Distributed Systems, Microcontrollers.
- 25 Socio-Humanities Cases: Criminal Jurisprudence & UNCLOS, Qualitative Management & Informant Saturation, Clinical Psychology & Likert Scales, Rural Sociology & Public Policy, Media Framing Discourse.
Universal Benchmark Results:
- ZeroGPT Live API: 130 / 130 (100.0%) achieved 0.0% AI.
-
GPTZero Statistical Engine: 130 / 130 (100.0%) achieved
HUMAN_ONLY(Mean Burstiness:$\mathbf{0.563}$ ).
This repository natively functions as an Agent Skill adhering to the Agent Skills Specification 1.0.
To install into your AI coding assistant (such as Google Antigravity, Claude Code, Cursor, or Roo-Code), place this directory in your skills path:
~/.gemini/config/skills/academic-writing-engine/
Your AI assistant will automatically recognize SKILL.md and apply academic prose restructuring and Word document generation autonomously.
Distributed under the MIT License. See LICENSE for details.
Developed by DemusBrain & Nicodemus (2026). Dedicated to advancing empirical scientific integrity and objective academic research.